Chenxu Wu, Qingpeng Kong, Peiang Zhao, Wendi Yang, Wenxin Ma, Fenghe Tang ,Zihang Jiang, S.Kevin Zhou

from MIRACLE Center, USTC

(Questions mail to 📧wuchenxu@mail.ustc.edu.cn)

Our method offers (a) superior quantitative performance, (b) improved qualitative results. It is (c) adaptable to various IR applications, (d) robust to different scales, and (e) resilient to different noise levels. $y$ represents the degraded image, $x_0$ denotes the sampling result, SR represents super-resolution and CS represents compressed-sensing.

To restore human face images, download this model(from SDEdit) and put it into DDNM/exp/logs/celeba/.

https://drive.google.com/file/d/1wSoA5fm_d6JBZk4RZ1SzWLMgev4WqH21/view?usp=share_link

To restore general images, download this model(from guided-diffusion) and put it into DDNM/exp/logs/imagenet/.

wget https://openaipublic.blob.core.windows.net/diffusion/jul-2021/256x256_diffusion_uncond.pt

Datasets can be accessed via the official repository of DDNM: DDNM GitHub Repository.

Download the CelebA testset and put it into DDNM/exp/datasets/celeba/.

Download the ImageNet testset and put it into DDNM/exp/datasets/imagenet/ and replace the file DDNM/exp/imagenet_val_1k.txt.

To execute EquS, kindly follow the instructions in the "evaluation.sh" script provided in the repository.

Our method remains equally effective with different image transformations:

(a,b) NFE vs. Evaluation metrics (block-based CS 25%). Our method is not limited by specific NFE. (c) Different transformations vs. Evaluation metrics. Random: Randomly select one transformation.

Please refer to our Paper for more results.

If you find this repository useful for your research, please cite the following work.

@article{wu2025equivariant,
  title={Equivariant Sampling for Improving Diffusion Model-based Image Restoration},
  author={Wu, Chenxu and Kong, Qingpeng and Zhao, Peiang and Yang, Wendi and Ma, Wenxin and Tang, Fenghe and Jiang, Zihang and Zhou, S Kevin},
  journal={arXiv preprint arXiv:2511.09965},
  year={2025}
}

This implementation is based on / inspired by:

• DDNM GitHub Repository

Thanks to the authors of DDNM for their great work.